PLUS: India ponders tax breaks for datacenters; Samsung plans hiring spree; Taliban bans fiber internet; and more
Asia In Brief Huawei last week revealed that China’s Zhejiang University used its Ascend 1000 accelerators to create a version of DeepSeek’s R1 model that improves on the original by producing fewer responses that China’s government would rather avoid.
The University explained how it used Huawei chips to build a post-training framework that re-worked DeepSeek’s model so it achieved “nearly 100 percent success rate against 14 common malicious threats, including toxic and harmful speech, politically sensitive content, and incitement to illegal behavior.” The re-trained model, dubbed DeepSeek-R1-Safe, is also more resistant to jailbreaking. The University achieved those results with just a one percent performance hit compared to DeepSeek-R1.
Huawei and the University are also proud to have outperformed the Alibaba Cloud’s Qwen model in terms of producing results that comply with China’s content rules.India considers 20-year datacenter tax holiday
India’s government last week reportedly discussed draft plans to give a 20-year tax holiday to developers who build new datacenters.
